Borrowed Confidence
My eldest is five. Today she walked the high rope at our local adventure playground, the one strung between two platforms about as high as the roof of our house. Halfway across she lifted a hand off the rope and waved at me, then carried on as though the whole thing were a walk to the corner shop.
The net is there. I read the safety notice on the way in, and she’s crossed that rope a dozen times while I watched from below. None of it stopped me feeling ill for every second she was up there.
The wave is older than the height. She didn’t learn it this afternoon. She learned it two years ago on the low ropes, the ones a foot above the bark chippings, where a fall means a graze and a short sulk. Hundreds of small drops down there trained her hands before her head could argue, so by the time she reached the high rope the balance was already paid for.
A good deal of what I build now arrives without that history. I’ll ask a model for a caching strategy, or a slab of infrastructure-as-code I didn’t write, and it comes back fluent and certain, waving from the top rope before I’ve glanced at the rigging. The assurance in the output is real enough. It’s simply not mine, and nothing in the field below it was ever bruised to earn it. I’m borrowing the confidence and shipping it under my own name.
This isn’t a complaint about the tools. They’re very good, I use them daily, and used well they make a team faster. But borrowed confidence changes what I should be looking at.
From the ground I can’t tell whether my daughter is genuinely that steady or one bad gust from the net. The walk looks identical in both cases. From a model, a wrong answer and a right one wear the same face: the generated migration that reverses cleanly on paper and locks the table under load reads just like the one that won’t. Fluency used to be a passable proxy for competence, because producing it cost something, and a person who wrote clearly had usually understood the problem first. That link has snapped, and it snaps hardest for people like me. Spend long enough reviewing other people’s work and you learn to treat fluency as a tell, to trust the code that reads well. That instinct was trained on a world where reading well and being right travelled together. A model gives you the first without the second, and aims it straight at the judgement I trust most. The research on model calibration bears this out, finding a stubborn gap between how confident these models sound and how often they’re right.
The net is where I look now.
The net is the unglamorous half of the job. A rollback I’ve actually executed, not one described in a runbook that no living engineer has run. The circuit breaker I’ve deliberately tripped, to watch the system catch something rather than assume it would. The net is whatever holds the fall I can no longer predict by admiring the walker.
Borrowed confidence makes that net the most important thing on the site. When I built slowly and by hand, my own hesitation was part of the safety system, the cheapest part I had, and it was never written down anywhere. So when generated work arrives without it, nothing flags the gap. The formal net, the tested recovery path, now has to carry the weight that instinct used to share. The reviewer’s judgement is still part of that net. The trap is letting the polish of the output pass for the scrutiny it hasn’t had.
All of which returns me to the foot of the rope, feeling ill. I used to treat that feeling as something to suppress, and I’ve changed my mind. The nerves are a signal, and a useful one: the readout from whoever is accountable for the catch. My daughter is free not to feel them, which is the privilege of being five. I’m not free, because I’m the one who checked the net was real.
Most of the attention, across the industry, goes to the walk. The demo is exciting and the velocity chart points the right way, while the recovery path earns a single line in a pull request and a quick approval. Good teams resist that on purpose, and the discipline counts for more now than it did three years ago: capable models produce more walks, and more convincing ones, which leaves far more rope to inspect and no extra hours in the day to do it.
The speed is welcome. The check simply has to move earlier. When work comes back fluent and ready to merge, the question worth your time is what you’ve actually watched it survive. Run the rollback yourself. Trip the breaker and watch it hold, before any of this reaches production. Confidence you did not earn is fine to lean on, so long as something you did earn is holding it up.
So I’ve stopped grading the wave. I grade the net.
She was down within the minute, off to the slide before I’d finished exhaling. Every year she gets steadier and I get no calmer. I’d rather keep it that way.
Reference
- Chhikara, P. (2025). Mind the Confidence Gap: Overconfidence, Calibration, and Distractor Effects in Large Language Models. Transactions on Machine Learning Research.
Views are my own and do not necessarily reflect those of my employer.